Attention-based Transformation from Latent Features to Point Clouds

This repository contains a PyTorch implementation of the paper:

Attention-based Transformation from Latent Features to Point Clouds
Kaiyi Zhang, Ximing Yang, Yuan Wu, Cheng Jin
AAAI 2022


In point cloud generation and completion, previous methods for transforming latent features to point clouds are generally based on fully connected layers (FC-based) or folding operations (Folding-based). However, point clouds generated by FC-based methods are usually troubled by outliers and rough surfaces. For folding-based methods, their data flow is large, convergence speed is slow, and they are also hard to handle the generation of non-smooth surfaces. In this work, we propose AXform, an attention-based method to transform latent features to point clouds. AXform first generates points in an interim space, using a fully connected layer. These interim points are then aggregated to generate the target point cloud. AXform takes both parameter sharing and data flow into account, which makes it has fewer outliers, fewer network parameters, and a faster convergence speed. The points generated by AXform do not have the strong 2-manifold constraint, which improves the generation of non-smooth surfaces. When AXform is expanded to multiple branches for local generations, the centripetal constraint makes it has properties of self-clustering and space consistency, which further enables unsupervised semantic segmentation. We also adopt this scheme and design AXformNet for point cloud completion. Considerable experiments on different datasets show that our methods achieve state-of-the-art results.


  • Python 3.6
  • CUDA 10.0
  • G++ or GCC 7.5
  • PyTorch. Codes are tested with version 1.6.0
  • (Optional) Visdom for visualization of the training process

Install all the following tools based on CUDA.

cd utils/furthestPointSampling
python3 install

cd utils/metrics/pytorch_structural_losses

cd utils/Pointnet2.PyTorch/pointnet2
python3 install

cd utils/PyTorchEMD
python3 install

# not used
cd utils/randPartial
python3 install


PCN dataset (Google Drive) are used for point cloud completion.

ShapeNetCore.v2.PC2048 (Google Drive) are used for the other tasks. The point clouds are uniformly sampled from the meshes in ShapeNetCore dataset (version 2). All the point clouds are centered and scaled to [-0.5, 0.5]. We follow the official split. The sample code based on PyTorch3D can be found in utils/

Please download them to the data directory.


All the arguments, e.g. gpu_ids, mode, method, hparas, num_branch, class_choice, visual, can be adjusted before training. For example:

# axform, airplane category, 16 branches
python3 --mode train --num_branch 16 --class_choice ['airplane']

# fc-based, car category
python3 models/ --mode train --method fc-based --class_choice ['car']

# l-gan, airplane category, not use axform
python3 models/latent_3d_points/ --mode train --method original --class_choice ['airplane'] --ae_ckpt_path path_to_ckpt_autoencoder.pth

# axformnet, all categories, integrated
python3 --mode train --method integrated --class_choice None

Pre-trained models

Here we provide pre-trained models (Google Drive) for point cloud completion. The following is the suggested way to evaluate the performance of the pre-trained models.

# vanilla
python3 --mode test --method vanilla --ckpt_path path_to_ckpt_vanilla.pth

# integrated
python3 --mode test --method integrated --ckpt_path path_to_ckpt_integrated.pth


Matplotlib is used for the visualization of results in the paper. Code for reference can be seen in utils/

Here we recommend using Mitsuba 2 for visualization. An example code can be found in Point Cloud Renderer.


Please cite our work if you find it useful:

 title={Attention-based Transformation from Latent Features to Point Clouds},
 author={Zhang, Kaiyi and Yang, Ximing, and Wu, Yuan and Jin, Cheng},
 journal={arXiv preprint arXiv:2112.05324},


This project Code is released under the MIT License (refer to the LICENSE file for details).


View Github